Prediction of Seoul House Price Index Using Deep Learning Algorithms with Multivariate Time Series Data
Tae Hyeong Lee, Myung‐Jin Jun · SH Urban Research & Insight · 2018
This study aims to evaluate the predictability of Deep Learning Neural Network algorithms (RNN and LSTM) in the forecast of the Seoul apartment price index. For the empirical analysis, we collect monthly housing price index data for the medium-sized and large-sized apartment units in Seoul during January 2006-October 2017 period. We also collect six macroeconomic variables that are known to affect housing price including expected inflation rate, rental price index, debt interest rate, stock price index, consumer price index, and unemployment rate. For the comparative purpose, we build Vector Autoregressive model (VAR) for multivariate time-series forecast. The analysis results indicate that the LSTM model best performed with the lowest RMSEs (0.826 and 1.038) for the medium-sized and large-sized apartment price indices, respectively, which is about 52 and 63 percent reductions from the VAR’s RMSE (1.708 and 2.825). We also found that standard deviations of predicted values from the LSTM are substantially lower than those of simple RNN, indicating higher stability of predicted price index from the LSTM than simple RNN.